Claude Code Weekly Limits Just Got +50% Through July 13. I Run Six Routines on Max. Here's the Codex Migration Math the Bump Doesn't Fix.

Anthropic raised Claude Code weekly caps by 50 percent through July 13, the third capacity bump in 27 days, and OpenAI countered the same afternoon with two free months of Codex for enterprise switchers. I run six publisher routines on Max 20x. The +50 buys us breathing room, not a fix.

What changed today, in three numbers

At 19:07 UTC today, @ClaudeDevs posted that weekly Claude Code limits are up 50 percent through July 13, 2026, live now for Pro, Max, Team, and seat-based Enterprise. The free plan is excluded. API users are untouched – they pay per token, with no weekly cap.

Reconstructed weekly ceilings after the bump, by tier: Pro at $20/mo lands around 60 Sonnet hours plus 12 Opus hours. Max 5x at $100 sits near 75 Opus hours. Max 20x at $200 – the tier I'm on – climbs from roughly 200 Opus hours to roughly 300. Team at $30 per seat shares the same percentage uplift across the pool.

Three minutes after the @ClaudeDevs post hit, @bridgemindai replied that the new caps put Claude Code "on par with Codex" and that they'd already cancelled Max twice over rate limits. Anthropic carefully did not engage. That non-engagement is the most informative part of the day. The official line is "more capacity for power users." The unofficial subtext is "please don't churn before July."

The three-bump sequence and what it actually says

This isn't one announcement. It's the third move in a 27-day defensive sequence, and the shape matters more than any single number.

Move one, April 16: the Claude 2x promo doubling off-peak capacity. A soft probe. Anthropic was measuring how much extra Opus consumption an unconstrained ceiling would draw, and whether the infra could carry it.

Move two, May 6: hourly Claude Code limits doubled, the peak-hours concept removed for Pro and Max, API rate limits up 1500 percent on input and 900 percent on output for Opus tier 1. Paired with a SpaceX Colossus 1 compute deal – 300 megawatts, 220,000+ NVIDIA GPUs – inside a single month. That's the supply side. You don't bolt on a third of a gigawatt of inference compute for a "nice to have."

Move three, today: +50 percent on weekly caps through July 13. The cap that hurt power users. The May 6 bump fixed hourly throttling, but cancellation tweets kept stacking because daily-limit relief without weekly-limit relief doesn't help anyone running real workloads for more than a Tuesday afternoon.

Splitting one capacity story into three press cycles is deliberate retention engineering. This is what an explicit anti-Codex defensive sequence looks like, paid for in 300 MW of fresh compute. The interpretation isn't subtle.

OpenAI's same-day counterattack and what the symmetry tells you

15:35 UTC, four hours before Anthropic posted: @OpenAIDevs wrote "Want to (officially) use Codex at work? Send this post to your CTO." Two free months of Codex usage for eligible enterprise customers who migrate inside 30 days.

Read the framing carefully. "Send this to your CTO" is not aimed at enthusiast developers. It's aimed at budget-owners. On a 50-developer Codex contract at roughly $30 per seat, two months equals $3,000 to $5,000 of real credit. On a 200- to 500-seat enterprise deal, it tops $18,000. That's a finance-department-sized incentive, not a developer perk.

The two moves are complementary, not duplicative. Anthropic defends the bottom of its funnel by lifting the ceiling for individual developers. OpenAI attacks the top by buying the CTO directly. Both companies are pressing the same migration pressure point from opposite ends.

Either way, the timing is interesting. It isn't comforting if you're Anthropic.

The token-per-task gap that does not move on July 13

Here is the number the +50 percent does not touch.

Independent benchmarks on identical workloads show Claude Code consumes 3 to 4x more tokens than Codex for equivalent output. Full Figma plugin build: Codex 1.499 million tokens, Claude Code 6.232 million – a 4.2x ratio. React plus Node scheduler app: Codex 72.5k, Claude Code 234.7k – 3.2x. REST API integration with OAuth: Codex roughly 180k, Claude Code roughly 650k – 3.6x.

Anthropic's narrative defense for this is "more deterministic and complete output." That defense does not work against GPT-5.3-Codex. SWE-Bench Pro is now a statistical tie – Codex 56.8 percent, Claude Code 55.4 percent. On Terminal-Bench 2.0 the gap inverts: GPT-5.3-Codex 77.3 percent, Claude Opus 4.6 65.4 percent. Twelve points in Codex's favor.

The economic consequence is brutal. At parity of subscription price, a user running heavy agentic workflows hits the Claude Code weekly cap in three to four days while closing the same task volume on Codex with margin to spare. The +50 percent narrows that gap. It does not close it. After July 13 the structural cost-per-task math reverts unless Anthropic intervenes on Opus token efficiency itself.

Two months of breathing room is what you buy when you can't ship the fix yet.

My six-routine stack: where the +50 actually changes the math

I run six publisher routines on top of Claude on Max 20x – each one a Cloudflare Workflow that picks, grounds, writes, cleans, persists, indexes, images, and publishes a category-specific article. Each step has a callAi cost chokepoint hardcoded to $1 per instance and $20 per day. I wrote about the production architecture in Cloudflare's 100x engineer stack decoded, and the cost discipline is what makes any of this affordable.

Pre-bump consumption, measured across the last six weeks: each publish cycle averages roughly 1.4 million Anthropic input tokens with prompt cache, plus 80k output. I run 4 to 6 dispatches a day across the six category routines. Weekly Opus time on the Max 20x dashboard hovered at 85 to 95 percent of the prior 200-hour cap. I was throttling routine #7 – a news-lane companion – because I couldn't safely add it without renegotiating Anthropic spend or porting half the chain to Sonnet.

Post-bump headroom, at the same dispatch rate: the cap lifted to roughly 300 weekly Opus hours moves consumption to 55 to 65 percent of weekly capacity. The news routine ships this week. I had it built and shelved.

The Codex comparison that remains honest: porting the same eight-step workflow to GPT-5.3-Codex would, by the 3 to 4x token ratio, cut consumption roughly 70 percent on identical output. The blocker isn't quality. It's that Codex's MCP and skills ecosystem is asymmetrically thinner than Claude Code's today, and the workflow steps where I need a first-attempt-correct refactor across 80 files still favor Opus 4.7 even at the higher token bill. The migration cost is real and load-bearing.

Net: the +50 makes Claude Code safe for my stack through July. It does not change the long answer. If by July 13 there's no Opus token-efficiency announcement, I'm scoping a partial Codex port for the routines where MCP coverage doesn't matter – image generation, indexing, persistence steps. Probably 40 percent of the consumption, ported.

For founders looking at the alternative-IDE side of this, the Cursor cloud agent environments piece covers what the parallel migration looks like outside the Codex/Claude binary.

What founders should actually do this week

If you're on Max 20x and were brushing the weekly cap: keep your seat. Push the workloads you've been throttling. You have until July 13, and the breathing room is real.

If you're on Pro and hit the weekly cap regularly: this is the cheapest upgrade window before the promo ends. Only worth it if your usage was consistently over 70 percent of the prior cap. If it wasn't, wait until July 14 to see what becomes permanent. Anthropic's historical pattern – including the Claude for Small Business announcement two weeks ago – is that promotional ceilings revert unless the underlying economics shifted.

If you have a CTO and a 50+ developer team currently on Claude Code: take the OpenAI two-month Codex credit seriously. The math is more honest after the +50 than before but still tilted in Codex's favor on token efficiency. The credit is worth $3K to $18K in real dollars, and the 30-day window forces a decision that you should be making anyway.

If you're uncertain: run both for two weeks on real workflows. Measure token-per-task on identical PRs. Decide on numbers, not on which CLI you find prettier. The /compact command, Caveman Mode, and PreToolUse hook tactics still cut 40 to 75 percent of Claude Code token consumption – apply those first before the comparison so you're benchmarking a tuned stack against Codex, not a default one.

The thing I'll be watching between now and July 13: whether Anthropic ships an Opus 4.7 or Opus 5 with materially better tokens-per-task, or whether they extend the promo. Extending the promo without efficiency gains is the tell. It would mean the underlying economics didn't move, and a second SpaceX-scale compute deal is the only way to sustain it. That's a thirty-day signal worth tracking.

Does the +50 percent weekly bump also apply to the free Claude Code plan?

No. The announcement explicitly names Pro, Max, Team, and seat-based Enterprise. The free plan is excluded.

Do I need to do anything to activate the new weekly limits?

No. The bump is automatic and live now for paying accounts. Verify in the Claude Code dashboard or by running /usage in the CLI.

Why is Anthropic raising limits if Opus is already winning on benchmarks?

It isn't. Independent benchmarks show Codex matches or beats Claude Code on SWE-Bench Pro and Terminal-Bench 2.0 while consuming 3 to 4x fewer tokens per task. The defensive bump targets weekly cap exhaustion, not benchmark perception.

Is this permanent, or does the cap revert on July 13?

Anthropic framed it as a time-bounded promo through July 13. Historical pattern – including the earlier spring Claude 2x promo – suggests caps revert at expiry unless another infrastructure-scale deal lands first.

Does the OpenAI counter-offer of two free Codex months apply to existing customers?

No. The 30-day Codex enterprise credit is explicitly for new enterprise customers migrating, not existing Codex teams.